A new VentureBeat Pulse Research survey points to a growing mismatch in enterprise AI: organizations are increasing infrastructure ambitions and considering provider changes even though many still lack basic visibility into what their existing AI compute actually costs.

The survey, based on 107 respondents from organizations with more than 100 employees in a single June 2026 wave, found that only 21% say they run AI in production at scale. Yet according to VentureBeat’s reporting, 45% plan to evaluate AI-specialized clouds over the next year, 64% intend to switch or add an infrastructure provider within 12 months, and 38% say that could happen within the next quarter. At the same time, 83% of respondents operating GPUs reported utilization at 50% or less, and only 44% said they rigorously track the cost and return of their AI compute.

That combination matters because it suggests the next phase of enterprise AI may be shaped less by raw demand for more hardware than by weak cost instrumentation, immature deployment practices, and hurried procurement decisions. For builders and enterprise buyers, the headline is not simply that AI budgets are rising. It is that spending appears to be moving ahead of operational discipline.

A market still early in deployment is already preparing to replatform

One of the clearest messages from the survey is that many companies remain early in their AI rollout. VentureBeat said 76% of respondents are either still experimenting or have only some workloads in production. That makes the current infrastructure reshuffle notable: organizations are not changing suppliers from a position of mature, stable operating experience. Many are still building the foundations.

Current usage remains concentrated around established providers. VentureBeat reported that Google Cloud led current AI infrastructure usage at 48%, with other major platforms including Microsoft, AWS and Oracle, plus model-provider APIs such as Gemini, OpenAI and Anthropic, making up most of the present-day stack. By contrast, specialized GPU cloud providers that attract heavy attention in AI infrastructure circles, including CoreWeave, Lambda, Crusoe and Nebius, registered little to no current use in this sample.

But planned evaluation looks very different from current deployment. The most-cited area for review over the next year was AI-specialized clouds, at 45%, according to the survey. VentureBeat also reported that 32% intend to evaluate non-Nvidia accelerators and 28% plan to evaluate next-generation Nvidia silicon. In other words, the next spending cycle may not deepen the status quo so much as test alternatives to it.

That is an important distinction for enterprise AI teams. A company can be heavily committed to cloud AI today while still using the next budgeting cycle to investigate new infrastructure layers, new accelerator types, and different cost structures for training and inference.

Buyers care about fit and TCO more than token sticker prices

The survey’s procurement signals are arguably more useful than its market-share snapshots. According to VentureBeat, respondents said integration with the existing stack was the top factor in selecting an AI infrastructure provider, cited by 41%, followed by total cost of ownership at 35%. Cost per million tokens ranked far lower, at just 8%.

That finding cuts against the popular framing of AI infrastructure competition as a race to post the cheapest unit price. Enterprise buyers, at least in this sample, appear to care more about whether new infrastructure fits identity systems, data pipelines, observability tooling, governance controls, and existing contracts than about a headline token rate.

The result also helps explain why Microsoft Azure and Google Cloud reportedly top near-term switching consideration at 33% each, with OpenAI at 30% and Gemini at 22%. VentureBeat interprets that as evidence that much of the near-term movement may involve incumbent providers trading share or consolidating customer spend, rather than a rapid migration to newer specialist vendors.

For product teams, that means technical merit alone may not win budget. Providers that can simplify migration, unify billing, improve governance, or make enterprise procurement easier may have more leverage than vendors focused narrowly on benchmark pricing.

The evidence points to an accountability problem, not just a capacity problem

The strongest statistic in the report may be the utilization data. VentureBeat said that among enterprises operating GPUs, 83% reported utilization at 50% or less, and 49% said utilization was 25% or below. Another 8% said they do not measure utilization at all.

If accurate, those numbers suggest a large amount of expensive capacity is sitting idle. That is not necessarily surprising in early-stage AI programs, where teams often overprovision for experimentation, buy capacity ahead of deployment, or lack workload schedulers sophisticated enough to keep systems busy. But it does undermine the idea that more GPUs alone are the immediate answer for most enterprise AI bottlenecks.

The accounting side looks similarly weak. Fewer than half of respondents, 44%, said they rigorously track compute cost and return, according to VentureBeat. Others said they track only partially, cannot quantify it yet, or have not prioritized it.

Taken together, those findings support the article’s core claim of a “compute gap”: enterprises are making faster and larger infrastructure decisions than their measurement systems can justify. In practical terms, that could mean teams approving reserved instances, GPU clusters, model API commitments, or new cloud partnerships without clear unit economics at the workload level.

For enterprise AI, this matters because cost problems usually appear late. A proof of concept can tolerate vague economics; a scaled inference product cannot. If organizations do not know which workloads are consuming budget, which models are overprovisioned, or where GPU queues and memory limits are creating hidden waste, the next phase of spending may simply make those inefficiencies harder to unwind.

Evidence, methodology and what is still uncertain

This is a useful survey, but it should be read cautiously. VentureBeat itself notes that the data comes from a single Q2 2026 wave, with 107 qualified respondents, and should be treated as a directional signal rather than a precise market measurement. The sample is self-selected, skews toward mid-market organizations, and leans toward earlier-stage AI adopters rather than the largest hyperscale operators.

That caveat matters for several reasons. First, current provider usage in this survey is not the same as global enterprise AI spending share. Second, reported switching intentions are often noisier than actual switching behavior. Third, utilization and cost-tracking responses are self-reported rather than independently audited.

Some of the more forward-looking claims are also survey-based attitudes, not observed market outcomes. The interest in AI-specialized clouds, non-Nvidia accelerators, and decentralized or sovereign compute reflects evaluation plans, not committed purchases. Likewise, concern about memory bandwidth and KV-cache constraints in inference is real in advanced deployments, but VentureBeat’s finding that roughly one in five respondents are unaware of the issue says more about buyer maturity than about when that bottleneck will become urgent for a given enterprise.

Even with those caveats, the survey’s internal logic is coherent: early-stage adopters are still building, they are dissatisfied enough to evaluate alternatives, and many do not yet have the cost controls needed to judge those alternatives well.

What this means for builders and enterprise AI teams

For infrastructure vendors, the message is clear: the enterprise sale is increasingly about operational visibility, not only access to accelerators. Providers that can show strong cost attribution, utilization reporting, policy controls, and workload orchestration may find more traction than those selling raw capacity alone.

For enterprises, the priority may need to shift from buying more compute to measuring current compute better. That means workload-level chargeback, model routing policies, GPU scheduling, caching strategies, and clearer split accounting between experimentation, training and inference. It also means comparing model API costs with self-hosted economics in a way that includes labor, integration, reliability, and governance overhead, not just token price.

For AI builders, especially those shipping products on Google Cloud, Microsoft Azure, AWS or Oracle while consuming APIs from OpenAI, Anthropic or Gemini, the survey is a reminder that infrastructure choices are no longer background decisions. They affect latency, margins, reliability, compliance, and product roadmap flexibility. If finance and engineering cannot jointly see those trade-offs, architecture decisions will be driven by procurement urgency rather than evidence.

What to watch next

The next signal to watch is whether stated evaluation interest turns into real workload movement toward CoreWeave, Lambda, Crusoe, Nebius or other AI-specialized clouds. A second key indicator is whether enterprises adopt more formal AI cost observability tools rather than continuing to manage spending through cloud dashboards and partial reporting.

It will also be worth watching whether the switching wave stays concentrated among incumbents such as Google Cloud, Microsoft Azure, OpenAI and Gemini, or broadens into meaningful accelerator diversification. If non-Nvidia evaluations convert into deployments, that would signal a deeper shift in how enterprises think about performance, supply risk and cost control.

Finally, watch the operational metrics. If future surveys show GPU utilization rising and more companies rigorously tracking compute economics, then today’s overspend may look like a temporary growing pain. If not, enterprise AI could enter its next expansion cycle with the same visibility problems, only on a larger bill.

Creati.ai perspective

The most important insight in this survey is not that enterprises want more AI infrastructure. That has been obvious for some time. The more useful signal is that many buyers now recognize integration and total cost of ownership as the real decision criteria, while still lacking the instrumentation to measure either with confidence.

That gap creates an opening across the stack. Cloud providers, model platforms, observability vendors and internal platform teams all have a chance to turn AI cost control into a product feature rather than a finance afterthought. The winners may be the companies that make compute legible: not just available, but attributable, governable and easy to compare across workloads. In enterprise AI, clearer economics may become a stronger competitive advantage than the next incremental drop in list price.